People now talk to AI assistants the way they talk to colleagues, and often more candidly. Agentic Collect, Right Forensics' AI conversation and agent collection service, collects those conversations the way we collect any chat between custodians, subjects and witnesses, then goes further. We reconstruct who asked what, what the assistant or agent did in response, including its searches and tool calls, and what the person did with the answer, and deliver it all in RSMF.
A text thread shows what someone chose to say to another person. A conversation with an AI assistant often shows something closer to how that person was thinking: the questions they could not answer themselves, the drafts they tried out, the shortcuts they considered.
That record can answer questions no other source reaches:
Same discipline as any chat collection. AI accounts are scoped, preserved, hashed and documented with the same care as phones, email and messaging apps, and placed on the same data-source map.
Structure kept, not flattened. AI conversations branch. A user can edit a prompt or regenerate a response, and an export may contain answers the user never saw on the final path, or only one path. We record branches, edits and which responses were on the path the user actually saw, where the export allows that to be determined.
Every line labeled by origin. Each entry is marked as a user prompt, an assistant response, a tool or search result, reasoning content, text extracted from an attachment, or a generated file. A reviewer can see who, or what, produced every line.
Hidden context surfaced. Custom instructions, saved memories and project files can shape responses without appearing in the chat. When the export carries them, we report them as a separate note on the conversation.
Limits stated plainly. Exports differ in which timestamps they keep, and conversations deleted before export generally do not appear, with no notice that anything is missing. We state those limits as properties of the source, not as findings about the person, and we never infer a message time without saying so.
Annotations kept apart from content. Our examiner notes travel inside the deliverable in clearly labeled fields, separate from the source text, so the fidelity of the conversation itself is never in question.
Every Agentic Collect delivery includes:
Details are generalized to protect the matters and the people involved.
The question. Counsel was investigating whether company information left with departing executives, and where it went afterward. The sources included work laptops, a personal phone, a thumb drive, personal webmail and personal accounts on consumer AI chat platforms.
What we did.
Why it mattered. The AI accounts were treated as first-class sources, not an afterthought. Because they were converted into the same review format as the rest of the messaging, counsel could follow a single timeline across devices, email and AI conversations.
The question. In a separate audit-related collection, messaging from three custodians' phones was produced in RSMF. After delivery, two different commercial processing tools were run on the same data as a quality check.
What we found. The two tools agreed exactly on the substance: the same conversations, dates and message counts. They differed sharply in presentation. One tool's custodian label inserted a common word into roughly seven of every ten documents, which would have produced thousands of false search hits. Other differences included time-zone handling that split some conversation-days across two local dates, encoding that broke words apart and defeated search, and summary counts in a tool's own logs that did not reconcile with its output.
Why it mattered. Because each of those issues was already measured and noted on the documents it affected, reviewers had the explanation in front of them, not a list of questions for later. We bring the same approach to AI conversations, which reviewers find even less familiar.
We work with major consumer and business AI assistants, including ChatGPT, Grok and Claude, and we assess others case by case. What can be preserved depends on what each platform's export and the custodian's devices actually contain.
Sometimes, partly. Platform exports generally leave out conversations deleted before the export, without saying so. Traces may remain on devices or in other sources. We will tell you what each source can and cannot show.
They can corroborate, but they show what a screen displayed, not what the account holds. They also miss branches, edits, attachments and tool activity.
Modern assistants do more than chat. They search the web, run tools, read uploaded files and produce documents. Agentic Collect captures those steps as the platform or device records them and labels each one, so you can tell what the person wrote, what the assistant wrote and what the assistant did.
RSMF lets your review team read AI conversations in the same platform, with the same search and coding tools, as the custodian's texts and chat apps.
AI conversations may hold the clearest record of what someone knew, intended or failed to check. Tell us about your matter and we'll map the sources and how to preserve them.
START A CONVERSATION